The presented work deals with the experimental identification of parts in a tree based decoder lexicon, that are more important for decoding efficiency compared to less important lexicon parts. Three different methods for constructing only the most important nodes in a set of tree lexicon copies are presented: building large trees; tree cutting; lexicon node removal. This leads to dramatic reduction of memory requirements while retaining the original recognition performance. In addition a reduction of the active decoding search space can be observed that leads to improved recognition speed. Although the presented methods can be generally applied to any HMM speech recognizer, experiments are performed in the hybrid MMI-connectionist/HMM system framework on the speaker independent 5k WSJ database.